P73Deep LearningAdvancedDiagnose14 min

Diagnose a Diverging Neural Network Training Run

Prioritize debugging from an exploding-loss trace.

#diagnose#deep learning#optimization

Scenario

Loss is 1.2, 1.0, 4.8, 91, NaN; gradient norms are 3, 8, 70, 900; one input is measured in millions and learning rate is 0.1.

Problem statement

Diagnose likely causes.

Your Task

  1. Read evidence.
  2. Prioritize fixes.
  3. Reject more epochs.

Related Concepts

Learning rateGradients

Ready to count this problem?

Mark it complete when you have reasoned through the solution in your own words.